Blog · arXiv Analysis · Published: August 12, 2026 · Modified: August 12, 2026 · Last reviewed: August 12, 2026

The Team Span Becomes the Sign-Off Boundary

A team can possess the actions needed to reach a goal while lacking anyone qualified to verify one of its requirements.

A new formal model separates possible motion, executable work, and valid completion. Its practical lesson is to roster the signers before the agents start.

The Paper

The source is Chainarong Amornbunchornvej’s Collective Counterfactual Planning: Coordination, Consent, and Verification under Representational Constraints, arXiv:2608.17932v1 [cs.MA, cross-listed cs.AI], submitted August 18, 2026. The 20-page version-one paper is affiliated with Thailand’s National Electronics and Computer Technology Center. It defines a formal planning problem and supplies a public reference implementation; it does not report a field study of human or AI teams.

Three Levels, Not One

The paper’s most useful move is to separate three meanings of solvable. Geometric feasibility asks whether available action directions could add up to the demanded change. Executable attainment asks whether agents can conceive and consent to a sequence that actually reaches the goal. Validated completion adds a final condition: every goal requirement must have a qualified verifier who can decide it from that verifier’s view. A reachable state, an executable plan, and an institutionally defensible finish are therefore different objects.

The Four Gates

In the Collective Counterfactual Planning model, agents occupy different subspaces of one task space. Four gates organize the work. Implementation names the coalition required for an action. Conception asks whether a prospective move looks like progress inside the proposing agent’s projection. Consent asks whether every required implementer finds the move sufficiently interpretable. Verification asks whether an agent’s projection preserves every distinction needed to certify a particular requirement. The first gate is assigned by the task; the other three are defined through the model’s representational geometry.

Relay Expands the Plan

The model’s positive result is relay. In its two-agent construction, one participant makes a move that changes a component visible to the other. That change makes a second move newly conceivable, even though no one-shot union of the agents’ initial plan trees contains a solution. The point is iterative reconsideration: publish an intermediate state, let every participant project it again, and continue. This is an existence result in a constructed geometry, not proof that adding agents or messages improves a deployed workflow.

The Dark Requirement

The negative result is sharper. The team span is the sum of everything represented by at least one agent; its orthogonal complement is dark to all of them. The paper’s unverifiability theorem says that if a requirement depends essentially on that dark subspace, no team member is qualified to certify it. The trajectory may accidentally land in the target state, yet the model will not permit a valid completion declaration. More effort cannot repair a missing basis for judgment.

For governance, that distinction is a warning about review theater. A planner, worker, critic, and judge can all be present while one requirement remains outside every role’s competence. Redundant votes do not create coverage when all voters share the same blind spot. The needed control is not another generic reviewer but a reviewer whose authorized evidence and expertise actually decide the uncovered requirement.

The paper also shows that consent cannot be summarized as a single strictness dial. Memoryless consent evaluates an action direction by itself; audited consent evaluates the cumulative state after the action. The formal examples establish that neither regime contains the other. A direction that looks opaque alone may be acceptable inside an interpretable accumulated state, while individually acceptable directions may accumulate into a state an implementer cannot endorse. In this model, consent is per required implementer and per step, not a vote on the whole plan.

Coverage Before Motion

The paper’s four-step exhaustive scheme begins with coverage preflight, then constructs the relay closure, ratifies candidate plans, and runs terminal inspection. That order deserves translation into ordinary governance: identify each requirement and a qualified signer before execution; record who may propose and perform each step; preserve intermediate states so other specialists can reconsider them; capture each required consent; and keep completion pending until every requirement has passed its own check.

For this essay, a sign-off-boundary receipt should name the goal requirements, action owners, required coalitions, role-specific evidence access, conception limits, consent rule, consent scope, verifier qualifications, uncovered requirements, intermediate-state log, completion test, search horizon, rejected plans, and correction route. The receipt should distinguish “the agents stopped,” “the state appears on target,” and “qualified checks establish every requirement.”

The Evidence Boundary

The paper deliberately assumes truthful agents, a shared goal, mechanical consent, public representational bases, fixed bases with no learning, and what it calls sovereign execution: a required participant can block a step it does not interpret. Its discussion says the mechanisms are face-valid rather than empirically validated. The computational figures use structured illustrations, and the practical solver chooses a restricted duration policy; a returned plan is sound under the formal predicates, but failure to find one is not a general impossibility certificate.

I inspected repository commit 9e066181 and ran its assertion suite. The relay, one-shot failure, two consent-separation examples, and adaptive-consent witness passed locally. That checks the deposited numerical examples against the deposited implementation; it does not independently prove the theorems or show that real organizations possess measurable vector subspaces of this kind.

The Spiralist lesson is modest: completion is a claim, not a light that turns on when an agent says “done.” Before a team is trusted to finish consequential work, the institution should prove that every requirement lies within somebody’s legitimate sign-off boundary.

Sources

Return to Blog